POET Built the Factory. NVIDIA Just Proved the Demand.

The optical-engine maker's new operations chief is racing to scale output as the AI buildout shifts toward the interconnect layer he supplies.

POET Technologies' chief operating officer has begun making the case that the company can mass-produce the optical engines feeding the artificial intelligence buildout. One day later, NVIDIA Corp. handed him the demand figures to back it up.

Dr. Sandeep Kumar, who took over as POET's operations chief on May 11, used his first public remarks to lay out a single proposition: the orders are coming, and the factory is ready. NVIDIA's quarterly results, released after the market closed on May 20, suggested the timing was no accident. The chipmaker's data center networking business — the layer of the AI stack where POET competes — nearly tripled from a year earlier.

The operations chief's pitch

Kumar arrived at $POET from Silicon Labs, where he ran worldwide operations for more than eighteen years. His brief on YouTube gets to the point and concrete: turn the company's facilities in Penang, Malaysia, into a high-volume production line. In his remarks, he said POET has built capacity to produce one million units a year of external light sources and optical engines, drawing on deposition, etching, and advanced packaging equipment deployed across its own operation centers and at its manufacturing partner, Nationgate.

Demand, he said, has grown extraordinarily over the past several quarters, a trend he tied to POET's optical interposer platform. The technology, in his telling, costs less, draws less power, takes up less space, and is built for the kind of volume manufacturing that AI infrastructure now requires.

On valuation, Kumar was deliberate rather than promotional. He framed any rerating as a consequence of execution, not a target in itself — the payoff for hitting customer delivery milestones, becoming a more predictable supplier, and keeping a motivated workforce. Rarely in his career, he said, had he seen a company with this kind of potential for high-volume growth.

What NVIDIA confirmed

NVIDIA's numbers gave the thesis a backbone. The company reported quarterly revenue of roughly $81.6 billion, up about 85% from a year earlier, with data center sales of some $75.2 billion, up 92%. It guided current-quarter revenue to around $91 billion, a sign the spending wave is still building. Chief Executive Officer Jensen Huang described the AI factory buildout as accelerating at extraordinary speed.

The figure that matters most for POET sits beneath the headline. NVIDIA's data center networking revenue jumped about 199% year over year, to roughly $14.8 billion, while its compute revenue grew 77%. The fastest-growing part of the AI data center, in other words, is no longer the processors — it is the plumbing that moves data between them. That is the territory of optical engines and light sources, and it is the bottleneck POET's pitch is designed to relieve: as racks grow denser and power budgets tighten, the constraint shifts from raw computing power to interconnect and efficiency.

The link, and its limits

The two stories rhyme. Kumar is describing the supply side of a demand wave that NVIDIA, the following evening, confirmed is accelerating fastest in exactly the layer POET addresses. POET's value proposition — cheaper, cooler, smaller, made at scale — maps onto the central problem of building AI data centers.

The technology underpinning that case, POET's Optical Interposer platform, is detailed on the company's technology page. A fuller treatment of the thesis appears in POET's Cannon Moment: A $500M Path Into the AI Stack, and the earnings read-through in NVIDIA Earnings Signal AI Data Center Ramp.

Two caveats temper the connection. NVIDIA's networking strength runs through its own NVLink, Spectrum, and Mellanox-derived products; it validates the optical-interconnect category, not POET's order book specifically. And Kumar's capacity and milestone claims remain forward-looking until they appear in reported revenue. His valuation language, in particular, was conditioned on execution — a framing worth preserving rather than recasting as a forecast.